PyOpenVINO - An Experimental Python Implementation of OpenVINO Inference Engine (minimum-set)

Description

The PyOpenVINO is a spin-off product from my deep learning algorithm study work. This project is aiming at neither practical performance nor rich functionalities. PyOpenVINO can load an OpenVINO IR model (.xml/.bin) and run it. The implementation is quite straightforward and naive. No Optimization technique is used. Thus, the code is easy to read and modify. Supported API is quite limited, but it mimics OpenVINO IE Python API. So, you can easily read and modify the sample code too.

Developed as a spin-off from my deep learning study work.
Very slow and limited functionality. Not a general DL inference engine.
Naive and straightforward code: (I hope) This is a good reference for learning deep-learning technology.
Extensible ops: Ops are implemented as plugins. You can easily add your ops as needed.

How to run

Steps 1 and 2 are optional since the converted MNIST IR model is provided.

(Optional) Train a model and generate a 'saved_model' with TensorFlow

python mnist-tf-training.py

The trained model data will be created under ./mnist-savedmodel directory.

(Optional) Convert TF saved_model into OpenVINO IR model
Prerequisite: You need to have OpenVINO installed (Model Optimizer is required).

convert-model.bat

Converted IR model (.xml/.bin) will be generated in ./models directory.

Run pyOpenVINO sample program

python test_pyopenvino.py

You'll see the output like this.

pyopenvino>python test_pyopenvino.py
inputs: [{'name': 'conv2d_input', 'type': 'Parameter', 'version': 'opset1', 'data': {'element_type': 'f32', 'shape': (1, 1, 28, 28)}, 'output': {0: {'precision': 'FP32', 'dims': (1, 1, 28, 28)}}}]
outputs: [{'name': 'Func/StatefulPartitionedCall/output/_11:0', 'type': 'Result', 'version': 'opset1', 'input': {0: {'precision': 'FP32', 'dims': (1, 10)}}}]
# node_name, time (sec)
conv2d_input Parameter, 0.0
conv2d_input/scale_copy Const, 0.0
StatefulPartitionedCall/sequential/conv2d/Conv2D Convolution, 0.11315417289733887
StatefulPartitionedCall/sequential/conv2d/BiasAdd/ReadVariableOp Const, 0.0
StatefulPartitionedCall/sequential/conv2d/BiasAdd/Add Add, 0.0
StatefulPartitionedCall/sequential/conv2d/Relu ReLU, 0.0010142326354980469
StatefulPartitionedCall/sequential/max_pooling2d/MaxPool MaxPool, 0.020931482315063477
StatefulPartitionedCall/sequential/conv2d_1/Conv2D/ReadVariableOp Const, 0.0
StatefulPartitionedCall/sequential/conv2d_1/Conv2D Convolution, 0.04333162307739258
StatefulPartitionedCall/sequential/conv2d_1/BiasAdd/ReadVariableOp Const, 0.0
StatefulPartitionedCall/sequential/conv2d_1/BiasAdd/Add Add, 0.0
StatefulPartitionedCall/sequential/conv2d_1/Relu ReLU, 0.0
StatefulPartitionedCall/sequential/max_pooling2d_1/MaxPool MaxPool, 0.006029367446899414
StatefulPartitionedCall/sequential/target_conv_layer/Conv2D/ReadVariableOp Const, 0.0010688304901123047
StatefulPartitionedCall/sequential/target_conv_layer/Conv2D Convolution, 0.004073381423950195
StatefulPartitionedCall/sequential/target_conv_layer/BiasAdd/ReadVariableOp Const, 0.0
StatefulPartitionedCall/sequential/target_conv_layer/BiasAdd/Add Add, 0.0
StatefulPartitionedCall/sequential/target_conv_layer/Relu ReLU, 0.0
StatefulPartitionedCall/sequential/target_conv_layer/Relu/Transpose/value6071024 Const, 0.0
StatefulPartitionedCall/sequential/target_conv_layer/Relu/Transpose Transpose, 0.0
StatefulPartitionedCall/sequential/flatten/Const Const, 0.0
StatefulPartitionedCall/sequential/flatten/Reshape Reshape, 0.0
StatefulPartitionedCall/sequential/dense/MatMul/ReadVariableOp Const, 0.0010004043579101562
StatefulPartitionedCall/sequential/dense/MatMul MatMul, 0.0013704299926757812
StatefulPartitionedCall/sequential/dense/BiasAdd/ReadVariableOp Const, 0.0
StatefulPartitionedCall/sequential/dense/BiasAdd/Add Add, 0.0
StatefulPartitionedCall/sequential/dense/Relu ReLU, 0.0
StatefulPartitionedCall/sequential/dense_1/MatMul/ReadVariableOp Const, 0.0
StatefulPartitionedCall/sequential/dense_1/MatMul MatMul, 0.0
StatefulPartitionedCall/sequential/dense_1/BiasAdd/ReadVariableOp Const, 0.0
StatefulPartitionedCall/sequential/dense_1/BiasAdd/Add Add, 0.0
StatefulPartitionedCall/sequential/dense_1/Softmax SoftMax, 0.0009992122650146484
Func/StatefulPartitionedCall/output/_11:0 Result, 0.0
@TOTAL_TIME, 0.21120882034301758
0.21120882034301758 sec/inf
Raw result: {'Func/StatefulPartitionedCall/output/_11:0': array([[7.8985136e-07, 2.0382247e-08, 9.9999917e-01, 1.0367385e-10,
        1.0184062e-10, 1.6024957e-12, 2.0729640e-10, 1.6014919e-08,
        6.5354638e-10, 9.5946295e-14]], dtype=float32)}
Result: [2 0 1 7 8 6 3 4 5 9]

Run Draw-and-Inter demo

python draw-and-infer.py

How to Operate

Left click to draw points.
Right click to clear the canvas.
This demo program is using 'numpy' kernels for performance.

A Littile Description of the Implementation

IR model internal representation

This inference engine uses networkx.DiGraph as the internal representation of the IR model. IR model will be translated into nodes and edges.
The nodes represent the ops, and it holds the attributes of the ops (e.g., strides, dilations, etc.).
The edges represent the connection between the nodes. The edges hold the port number for both ends.
The intermediate output from the nodes (feature maps) will be stored in the data attributes in the output port of the node (G.nodes[node_id_num]['output'][port_num]['data'] = feat_map)

An example of the contents (attributes) of a node

node id= 14
 name : StatefulPartitionedCall/sequential/target_conv_layer/Conv2D
 type : Convolution
 version : opset1
 data :
     auto_pad : valid
     dilations : 1, 1
     pads_begin : 0, 0
     pads_end : 0, 0
     strides : 1, 1
 input :
     0 :
         precision : FP32
         dims : (1, 64, 5, 5)
     1 :
         precision : FP32
         dims : (64, 64, 3, 3)
 output :
     2 :
         precision : FP32
         dims : (1, 64, 3, 3)

An example of the contents of an edge

format = (from-layer, from-port, to-layer, to-port)

edge_id= (0, 2)
   {'connection': (0, 0, 2, 0)}

Ops plugins

Operators are implemented as plugins. You can develop an Op in Python and place the file in the op_plugins directory. The inference_engine of pyOpenVINO will search the Python source files in the op_plugins directory at the start time and register them as the Ops plugin.
The file name of the Ops plugin will be treated as the Op name, so it must match the layer type attribute field in the IR XML file.
The inference engine will call the compute() function of the plugin to perform the calculation. The compute() function is the only API between the inference engine and the plugin. The inference engine will collect the required input data and pass it to the compute() function. The input data is in the form of Python dict. ({port_num:data[, port_num:data[, ...]]})
The op needs to calculate the result from the input data and return it as a Python dict. ({port_num:result[, port_num:result[, ...]]})

Kernel implementation: NumPy version and Naive version

Not all, but some Ops have dual kernel implementation, a naive implementation (easy to read), and a NumPy version implementation (a bit faster).
The NumPy version might be x10+ faster than the naive version.
The kernel type can be specified with Executable_Network.kernel_type attribute. You can specify eitgher one of 'naive' (default) or 'numpy'. Please refer to the sample program test_pyopenvino.py for the details.

END

Experimental Python implementation of OpenVINO Inference Engine (very slow, limited functionality). All codes are written in Python. Easy to read and modify.

Related tags

Overview

PyOpenVINO - An Experimental Python Implementation of OpenVINO Inference Engine (minimum-set)

Description

How to run

How to Operate

A Littile Description of the Implementation

IR model internal representation

An example of the contents (attributes) of a node

An example of the contents of an edge

Ops plugins

Kernel implementation: NumPy version and Naive version

Owner

Yasunori Shimura

Implementation of Artificial Neural Network Algorithm

This is the workbook I created while I was studying for the Qiskit Associate Developer exam. I hope this becomes useful to others as it was for me :)

Cross-modal Retrieval using Transformer Encoder Reasoning Networks (TERN). With use of Metric Learning and FAISS for fast similarity search on GPU

(ICCV 2021) ProHMR - Probabilistic Modeling for Human Mesh Recovery

An Agnostic Computer Vision Framework - Pluggable to any Training Library: Fastai, Pytorch-Lightning with more to come

Normal Learning in Videos with Attention Prototype Network

Geometric Sensitivity Decomposition

A simple rest api serving a deep learning model that classifies human gender based on their faces. (vgg16 transfare learning)

Improved Fitness Optimization Landscapes for Sequence Design

This is a tensorflow-based rotation detection benchmark, also called AlphaRotate.

Multiple custom object count and detection using YOLOv3-Tiny method

The code is an implementation of Feedback Convolutional Neural Network for Visual Localization and Segmentation.

[CVPR 2021] NormalFusion: Real-Time Acquisition of Surface Normals for High-Resolution RGB-D Scanning

tree-math: mathematical operations for JAX pytrees

Novel Instances Mining with Pseudo-Margin Evaluation for Few-Shot Object Detection

Tianshou - An elegant PyTorch deep reinforcement learning library.

Multi-scale discriminator feature-wise loss function

Network Enhancement implementation in pytorch

TGS Salt Identification Challenge

Author's PyTorch implementation of Randomized Ensembled Double Q-Learning (REDQ) algorithm.